Ishowbackup/gemma-4-12B-it-uncensored

TEXT GENERATIONPricing:Input $1.2 / Cached $0.24 / Output $4.8Concurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 18, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Ishowbackup/gemma-4-12B-it-uncensored is a 12 billion parameter instruction-tuned causal language model based on Google's Gemma4Unified architecture, with a 32768 token context length. This model has been specifically modified to significantly reduce refusal behavior, achieving a 2.0% refusal rate across multiple datasets compared to the original model's high refusal rates. It utilizes a norm-preserving biprojected abliteration method to remove refusal signals while maintaining response quality, making it suitable for applications requiring less restrictive content generation.

Loading preview...

Overview

Ishowbackup/gemma-4-12B-it-uncensored is a 12 billion parameter instruction-tuned model derived from Google's gemma-4-12B-it (Gemma4Unified architecture). Its primary distinction is the removal of refusal behavior, making it an "uncensored" variant. This modification was achieved using a novel norm-preserving biprojected abliteration method, which projects out refusal directions from the model's weights without degrading overall response quality.

Key Capabilities & Differentiators

  • Significantly Reduced Refusals: Achieves a refusal rate of 2.0% across a cross-dataset validation of 686 prompts, a substantial reduction from the original model's near-100% refusal rate on certain harmful prompts.
  • Quality Preservation: Audits and Q8 inference verified no degradation in response coherence or quality despite the refusal removal.
  • Advanced Abliteration Method: Employs a norm-preserving biprojection technique that maintains weight magnitudes, uses per-layer refusal directions, and is a deterministic single-pass process, differing from standard projection methods.
  • Targeted Modification: Refusal signals were concentrated and abliterated in the upper 70% of decoder layers (L15-47) of the Gemma4Unified architecture.

Use Cases

This model is ideal for applications where a less restrictive language model is desired, particularly for tasks that might otherwise trigger refusal behaviors in heavily moderated models. Developers can leverage its uncensored nature for creative writing, role-playing, or research into model safety and bias without encountering frequent content restrictions.

Finetunes
2 models